I don't think anyone is talking about it because it's not a very productive conversation to have. I'm not particularly bullish on vibe coding either but if you could explain what exactly about vibe coding causes these specific issues then it could be more interesting to discuss.
But as it stands, the more likely reason is capacity crunch caused by a chips shortage and demand heavily outpacing supply. You vibe coding reason is based on as much vibes as their code probably is.
Vibe coding does not usually produce performant code, it produces spaghetti with the goal of making the user asking for work to be done to go away as soon as possible with a (often barely) working solution.
I recently vibe-translated a simple project from Javascript to C, where Javascript was producing 30fps, and the first C version produced 1 frame every 20 seconds. After some time trying to get the AI to optimize it, I arrived at 1fps from the C project. Not a win, but the AI did produce working C code.
I have no doubt that if I had done this myself (which I will do soon), with the appropriate level of care, it would have been 30fps or more.
that is a separate issue indeed, but their comms make it rather obvious they are scrambling to reduce compute and they're just slashing their service selectively - with openclaw and max users being the first in the chopping block
That's not an elephant in the room.. it's just proof of how insanely useful the tool is and the reality that so much more hardware is needed. Thus people saying "why are these companies building insanely large data centers" ... this is why!
The problem is that vibe-coding, when it fails (i.e. it's non-useful, at least for a bit), is usually solved by more vibes. Try again and hope it works. Ask it to refactor and hope the cleaner code helps it along. If you're willing to think about the code yourself you'll likely ask it questions about the codebase. High vibe-code usage is both a metric that it is working and that it's failing.
I think it is telling that this audience down votes this. It's kind of obvious that the thing is being used a lot. Doesn't mean it works as well as advertised. Doesn't mean the business model they have works. Just means there is a lot of demand. You can't ignore that.
I have no particular insight into the Anthropic backend, but it's possible in general for systems to have architectural issues which cannot be mitigated by just adding more hardware.
The proof is already there. It's concrete. I've seen it directly the last few months of using claude code. It closed the loop. It's insanely beneficial when used properly- that is a pure fact. You act like it's an opinion.
If an AI doesnt generate perfect code, if left to its devices it will at some point create a codebase big and nasty enough that it will not be able to deal with it.
It should catch up faster. It's absolutely useless for the bulk of the tedium—notably, soldering together random repos to satisfy executives—that makes up my job now.
Google's customer support is interesting. Its definitely a case where you'll sometimes hit pockets of the company where clearly there was someone who made it their life's work to fix this bad reputation they have; while other pockets make it clear that they deserve the reputation.
I had a Nest subscription that became a total mess. If you've ever tried to use Nest before, or are coming from a legacy Nest account, and/or also have a Workspace account that somehow got wrapped up in the mess, you'll understand how much of a clusterf Nest is for the Google ecosystem. I had signed up for this subscription on a personal Google account, cancelled it, but was still being charged for it, and the credit card being used made me think it was getting charged on my Google Workspace account (which isn't officially supported, and would never let you sign up for it, but DID share an email address with my legacy Nest account I had migrated into the non-Workspace personal account I was using for Nest).
They had to escalate the problem a couple times, which took ~24 hours. Once that happened, their rep had it resolved in minutes, and refunded me two months on the subscription.
The biggest piece of advice I can give when dealing with Google is: Never be weird. You cannot ever put yourself in a situation where your account isn't like the other billion accounts they have. If you do, something will go wrong and its rolling the dice on whether you'll ever reach someone who can help you. If you've used Google enough, you know: Their multifactor settings are weird. You cannot set it up exactly how you want; it'll always trigger some auth method you didn't configure but they have "LATENT KNOWLEDGE" you should be able to authenticate with, like a phone number you configured six years ago, or gmail installed on a tablet that's 400 miles away, and you can't turn it off, even on Workspace.
My favorite bit of Googleism: Go to any site you sign in with Google SSO and watch the URLs in the eight redirects it has to do before it signs you in. You'll see a "youtube.com" in there. Even on a Workspace account. Youtube.com is a load-bearing website in their core auth flows.
Mess of a company. I hope they invest some effort in improving things, but I was saying the same thing in 2018. They probably won't.
> like a phone number you configured six years ago
I've put in a heroic effort to make sure they never get a phone number, specifically so they can't start handing my account over to the first clown who simswaps me, and have been successful. Unfortunately, this makes my account weird, which as you noted is fatal.
> My favorite bit of Googleism: Go to any site you sign in with Google SSO and watch the URLs in the eight redirects it has to do before it signs you in. You'll see a "youtube.com" in there. Even on a Workspace account. Youtube.com is a load-bearing website in their core auth flows.
I assume that's just because they need to set a cookie on the YouTube domain in case you visit YouTube later on the workspace account, and not "load bearing"in the manner you insinuate
You do not hit YouTube directly. You hit Google's frontend server which then does internal routing. Likely it would be able to route around it. Or rather, the auth part of YouTube is not the same as all of the other parts of YouTube. For big tech companies like Google, a website is not one single binary that serves requests, but a ton of services handling different things (and some of those services being caching so it doesn't show things down as you might think). It is highly unlikely that the main services comprising YouTube as a video streaming site would bring down Google auth
We need to come to terms with the possibly-irreparable harm that private capital has done to the West. Capital is long-past serving the interests of the broader public; but we're now past the point of capital serving the interests of the corporations its being invested into. The demand for shares in OpenAI and Anthropic is so high that its pushing their valuations into territories they can never hope to drive revenue to fulfill; the cycle of this massive warchest of private capital inside the AI industry has for all intents and purposes created a communist economic structure, with all of its faults. Grifting, favored suppliers; if the stories about SK Hynix and Samsung guaranteeing 40% of their wafer supply to OpenAI on a letter of intent they cannot follow through on are true, we're even getting good old fashioned communist mis-allocation of resources. The day may eventually come when the USG is forced with the decision of bailing out the trillion dollar OpenAI Corporation; taking a stake to add to their portfolio next to Intel and others; and maybe normies will then realize what is happening, but the writing has been on the wall for years.
I love capitalism; its ability to allocate resources on a macroeconomic scale to pick winners and more-importantly losers doesn't have a rival system. As a younger, more naive startup employee, I'm on the record making a total fool of myself responding to our CEO talking about struggling to find PMF by saying "then maybe our company doesn't deserve to exist" (yeah...) But the "capitalists" who run the world aren't actually interested in capitalism, and thus definitionally can't be capitalists. At least once upon a time we had filthy rich titans you could look up to, like Buffet and Gates (Epstein stuff aside); but at this point most of them aren't even enviable people. Despite being richer than God, people like Huang, Musk, Ellison, and Zuckerberg feel more like vampires; they want to spend their whole lives doing the exact same thing, getting richer and richer, refusing to put a ladder down for anyone else to take a shot at improving on what they've built. I actually have a modicum of respect for Bezos and, to whatever vanishingly small degree, Trump; at least they're trying something different.
The wording is weird enough that I have to agree. This is the first time I've ever heard spyware segmented using "mercenary" as a qualifier, which is just insanely suspicious.
Oh geez. Legal did not give them the go ahead to make the unqualified statement: “We are not aware of any successful spyware attacks” they had to explicitly qualify it with “mercenary”.
There are more weasel words "we are not aware" - means they actually don't know if such attack was successful, "successful" - what is the definition of success? Maybe attackers got access, but didn't find anything interesting?
I think you are, the words make perfect sense. They know of a lot of attack attempts, and so far they have no reason to believe any were successful. Success can mean a lot of different things, why list it all out (were able to extract data, install malicious software, encrypt files with ransomware, delete any data, etc).
They have a legal department carefully directing what they say. In a court of law, their lawyers will successfully argue that they are beholden to only the precise letter of their statement. Are you arguing that their lawyers are incompetent and imprecise in their wording? If so, what evidence do you have that their lawyers are incompetent?
In light of the correct legal interpretation of their words, being only the specific letters, we can see that your interpretation is incorrect.
> They know of a lot of attack attempts
No, their statement says nothing about attack attempts.
> so far they have no reason to believe any were successful
No, their statement says nothing about their belief, only their explicit knowledge. Their statement says nothing about their investigation practices or whether they even attempted to investigate and learn about attacks. Their statement says nothing about non-mercenary attacks.
Their statement is technically correct as long as any successful attacks they know about are not explicitly known to be committed by mercenarys.
> No, their statement says nothing about attack attempts.
That's a good point. The best way not to know about any successful attacks is not to know about any of them. I also can definitively state that I'm not aware of any successful attacks, but for obvious reasons this is a basically meaningless statement. Without more data, it's not clear how meaningful the statement they gave is, and while it probably is more meaningful than mine, it doesn't make sense to jump from what they said to "there have definitively been no successful attacks" based on it.
I'm just going to ignore your entire first paragraph that tries to use hostility to overcome a clear willful misunderstanding, or strong evidence of a recent stroke.
> No, their statement says nothing about attack attempts.
Exactly, they're keeping the statement brief and correct. They have sent multiple batches of notifications to users on previous attacks.
The statement is clear, covers their primary use case for the product, and I'm sure is legally sound. You're grasping at straws trying to think up ways they can be lying to you. I would be very surprised if you ever have used their lockdown mode with any actual cause.
I am glad that you agree that their legal department’s explicit and intentional exclusion of known successful non-mercenary attacks is precise and legally sound.
It is advisable to not grasp at straws to think up ways that highly paid lawyers are not saying exactly the words they have approved. That is literally their job and they are good at it.
If they meant something more expansive they can do so. It is not the public’s job to do it for them while letting them retreat to the legally binding interpretation at their pleasure.
They can be perfectly aware of nation-state hacks. These are exactly the weasel qualifiers used by the NSA when they were claiming not to be watching the communications of US citizens. "No intercepts were made under program X" specifically sidesteps all the shady stuff under program Y.
They have very good reason to believe that - shareholders and public perception. Apple maintains image of their phone being secure and that is far from the truth. As long as general public don't know their phones have holes like Swiss cheese, the shareholders will be happy.
>"successful" - what is the definition of success?
At risk of stating the obvious, isn't success "hacked it and no one ever found out (at the time)"? By definition, Apple could probably only be aware of unsuccessful attacks. Though that's not guaranteed either, considering all the myriad failure modes that there must be.
I don't want to give too much credit to Github, because their uptime is truly horrendous and they need to fix it. But: I've felt like its a little unfair to judge the uptime of company platforms like this; by saying "if any feature at all is down, its all down" and then translating that into 9s for the platform.
I never use Github Copilot; it does go down a lot, if their status page is to be believed; I don't really care when it goes down, because it going down doesn't bring down the rest of Github. I care about Github's uptime ignoring Copilot. Everyone's slice of what they care about is a little different, so the only correct way to speak on Github's uptime is to be precise and probably focus on a lot of the core stuff that tons of people care about and that's been struggling lately: Core git operations, website functionality, api access, actions, etc.
> I've felt like its a little unfair to judge the uptime of company platforms like this; by saying "if any feature at all is down, its all down" and then translating that into 9s for the platform.
This is definitely true.
At the same time, none of the individual services has hit 3x9 uptime in the last 90 days [0], which is their Enterprise SLA [1] ...
> "Uptime" is the percentage of total possible minutes the applicable GitHub service was available in a given calendar quarter. GitHub commits to maintain at least 99.9% Uptime for the applicable GitHub service.
> If GitHub does not meet the SLA, Customer will be entitled to service credit to Customer's account ("Service Credits") based on the calculation below ("Service Credits Calculation").
The linked document in my previous comment has more detail.
It's worth adding that big (BIG!) business clients will usually negotiate the terms for going below the SLA threshold. The goal is less to be compensated if it happens, and more to incentivize the provider to never let it happen.
Right. Basically, they give you a coupon to lower your cost of future consumption. So, you have to keep consuming the service. If you just leave, you get no rebate. Obviously, very large customers get special deals.
You're right that labelling any outage as "Github is down" is an overgeneralisation, & we should focus on bottlenecks that impact teams in a time sensitive matter, but that isn't the case here. Their most stable service (API) has only two 9s (99.69%).
They're not even struggling to get their average to three 9s, they're struggling to get ANY service to three 9s. They're struggling to get many services to two 9s.
Copilot may be the least stable at one 9, but the services I would consider most critical (Git & Actions) are also at one 9.
Most people complaining about uptime aren't free users or open-source developers. It's people whose companies are enterprise GitHub customers. It's a real problem and affects productivity.
The issue is also that those 27 hours don't happen at once, They happen in small chunks of a couple minutes which makes it happen almost everyday and has a ton of downstream build and retry issues. The resulting downtime is probably 2 orders of magnitudes higher at least.
Those 27 hours only seem to happen during the workday when I’m trying to push branches, run CI pipelines or otherwise use GitHub (I don’t use Copilot). Whatever the yearly figure, it’s been a pain in the ass these last few months and it’s unacceptable, free or no (my company pays for GitHub).
Honestly, you're right - 2̶7̵ 87+ (correction from sibling) hours per year is absolutely fine & normal for me & anything I want to run. I personally think it should be fine for everybody.
On the other hand the baseline minimal Github Enterprise plan with no features (no Copilot, GHAS, etc.) runs a medium sized company $1m+ per annum, not including pay-per-use extras like CI minutes. As an individual I'm not the target audience for that invoice, but I can envisage whomever is wanting a couple of 9s to go with it. As a treat.
This company is part of the portfolio of a $trillion+ transnational corporation. The idea that we can't judge them, when they clearly have more resources than 99% of other companies on this planet, doesn't hold up to any scrutiny.
Why defend a company that clearly doesn't care about its customers and see them as a money spigot to suck dry?
The OP clearly never says we can't judge them. He was speaking to how the uptime is measured. I'm not saying I agree or disgree with the OP but at least address the argument he's making.
It doesn't help that almost all of the big tech companies talking about 5 9s are lying about it; "Does it respond to the API at all, even with errors? It's up!" and so on. If you spend a lot of time analyzing browser traces you see errors and failures constantly from everyone, even huge companies that brag a lot about their prowess. But it's "up" even if a shard is completely down.
The five nines tech people usually are talking about is a fiction; the only place where the measure is really real is in networking, specifically service provider networking, otherwise it's often just various ways of cleverly slicing the data to keep the status screen green. A dead giveaway is a gander at the SLAs and all the ways the SLAs are basically worthless for almost everyone in the space.
See also all of the "1 hour response time" SLAs from open source wrapper companies. Yes, in one hour they will create a case and give you case ID. But that's not how they describe it.
Once you dig into the details what does it mean to have 5 9s? Some systems have a huge surface area of calls and views. If the main web page is down but the entire backend API still is responding fine is that a 'down'? Well sorta. Or what if one misc API that some users only call during onboarding is down does that count? Well technically yes.
It depends on your users and what path they use and what is the general path.
Then add in response times to those down items. Those are usually made up too.
I think there's a reason why the people from Moonshot deleted their tweets; they're probably just researchers who got yelled at by the people who actually knew what was going on at Moonshot.
People need to seriously stop it with the whole reddit-esque Boston Marathon Bomber investigation-style low-info crusades. Its extremely unhealthy for both your own mental state and the state of discourse on the internet. Even if Cursor misbehaved (they did not): Your life is not materially changed whether they did or did not. Use it, or don't use it; these things are a matter that lies exclusively between Cursor and Moonshot.
Can you ensure that Notion is able to keep delivering given they don't develop their own models? Lovable? OpenCode? Should we be worried that Discord might disappear because they don't run their own data centers? Personally, I'm very concerned that one day Google might just have to close up shop, because while they do design their own chips, they don't fabricate them in-house; and don't get me started on TSMC and their critical dependency on ASML, they might as well just lock the doors.
This is exactly what Cursor should be doing, within the obvious bounds of the law and such. Not everyone needs a pristine foundation model. What a waste of compute. Anthropic & OpenAI need product-level competition to knock them off their $25/Mtok horse.
I'm really struggling to understand how Anthropic is benefited by not allowing this. Its bad PR for no good reason. The only thing I can figure is that Claude Code is hemorrhaging money, they're too afraid to actually enforce reasonable token limits, and the only thing that's keeping it from totally bankrupting the company tomorrow is: controlling the harness and having the harness dynamically route toward Haiku or Sonnet over Opus when Opus is overloaded, without telling the user. Or maybe, they're extremely interested in observability of the exact prompts users are typing, and third party harnesses muck that data in with the rest of the context that gets sent, so its harder to detangle the prompt from the noise?
Like, in any event, I seriously get the feeling that Anthropic doesn't just not care about their users, but actively dislikes them. Like, they must be losing so much money on each Claude Code subscriber that if a million people all said "we're switching" they just wouldn't care. I get this vibe even from watching videos of people working on the Antrhopic team; like they all think they're Gods above mere mortals, serving some higher purpose, and nothing matters to them except Building the Machine God.
OpenCode is awesome. Claude Code is nothing special at all. Last month I switched to just using OpenCode with a Codex $200/mo subscription, and that's been great. Let the weirdos at Anthropic do what weirdos do, and hopefully one day their name is never mentioned again in polite society.
This is analogous to when Google launched Gmail with 1GB of storage and then a bunch of third-party apps cropped up that took advantage of it to use it as a generic online file storage drive.
There was GMailFS[0] and Gmail Drive[1] - this is before S3, dropbox, and a time where web hosting would give you ~10MB or so of space.
Google updated their ToS and shut down accounts using their service in ways they weren't intended via these apps - because obviously the 1GB of storage was a loss-leader into Google's ecosystem (and it worked)
Same thing today - "unauthorized" third parties taking advantage of a loss-leading[2] deal - complete with similar trademark violations to boot[3].
Google have more cash to burn in the AI race so can be more forgiving today in how their codex plans are used. Anthropic are still a private company and can't.
[2] it's a big q just how large a loss leader the max plans are considering a fixed harness, prompt caching etc. but point still stands. you're getting up to $5k of RRP tokens for $200
No, because in those cases you're still a user of gmail. When you tell people your email address, or send people email, and it contains "@gmail.com", you're still implicitly advertising for Google. From Google's perspective that's still worth the few KB per day of bandwidth and 1GB storage (which the vast majority of people never use the entirety of, anyway) they're giving away.
But when you use gmail accounts as file storage, you're both a higher-cost user and also doing nothing to further Google's ecosystem (since the email address itself is probably not being used for genuine messaging at all).
And here, you're still using Claude Opus, and when people ask you what you used, you'd say OpenCode client with Claude (Thunderbird client with Gmail).
there is nothing about claude code that prevents you from using it for non coding use cases. nothing that happens in open code or any harness for that matter is hidden from anthropic. neither does open code allow access in some nefarious use case that claude code does not.
the difference is not like the difference between gmail and gmailfs like you seem to be misunderstanding. a more accurate comparison would be the difference between curl, or httpie vs postman.
It's not analogous at all because Google intentionally provided interfaces for those clients and even instructions for using them.
An analogous situation would be if someone reverse engineered the Google Maps API and provided their own app that showed maps using the Google Maps data.
> And if Google Maps charged per tile viewed, so the user pays the same amount regardless of which maps client they used, would your opinion hold?
Yes. Why wouldn't it hold?
Anthropic has a pay-per-token API. You can use OpenCode with it.
Maybe my consistency comes from having worked with contracts and agreements in the real world, where the end user doesn't get to pick and choose which terms they want to abide by.
When you sign up to use a service, you're not signing up to use it however you would like, on your own terms. You're paying for a service that they offer. They are not obligated to continue offering it to you if you try to use it a different way.
My point is that model providers are just a compute service, and should have no say in what sends the data, or displays the data. Especially when they only bill based on the quantity of data.
They have an API for exactly that. You can use it.
They offer a separate plan with discounts for use with their tools. You can also choose to take advantage of those discounts with the monthly fee, within the domain where that applies. You cannot, however, expect to demand that discount to apply to anything you want.
You can argue about what you want it to be all day long, but when you go to the subscription page and choose what to purchase it's very clear what you're getting.
> They are basically a utility
Utilities like my electric company also have different plans for different uses. I cannot, for example, sign up for a residential plan and then try to connect it to my commercial business, even though I'm consuming power from them either way.
Utilities do not work like that. They do have contractual agreements about how you can use the resources provided.
> Google have more cash to burn in the AI race so can be more forgiving today in how their codex plans are used.
Even despite the larger cash pile to burn, Google is in the middle of their own controversy around what many feel is a rug-pull around how Gemini "AI credits" work and are priced.
This argument is predicated on Anthropic losing money on the subs, but I'm not sure that's a cut and dried argument. OpenAI have said publicly that they're (very) profitable on inference, and they're much cheaper than Anthropic. I suspect this is just artificially trying to create a moat. The problem is their moat is not as sticky as they think it is - I completely ditched Claude for Codex a while ago, my money now goes to OpenAI, and I'm very happy with it. For a while Claude was noticeably better, but that's not the case any more - in my case I prefer Codex.
They aren't public companies (yet). They are allowed to just lie about these things. It's also not really reasonable to only count inference compute as a cost since it's not like any of these companies could stop doing R&D without being abandoned for having worse models within a year or 2
Here's one possibility: Anthropic understands the value of the brand and the harness and that those two things are connected, specifically because they came from behind. OpenAI almost accidentally launched a global brand overnight. ChatGPT went from nothing to the kind of english word you hear in non-anglophone countries in about a month. Millions and millions have used it (at least once) and more people associate it with AI than use it. OpenAI's problem is managing the big industry links so that by the time the hype cools down, they're already plugged into tools. Their "moat" is that number of companies that matter is actually small and all those companies like predictable, enterprise shaped solutions with contracts and stuff. Unlike developers who might switch their subscriptions quickly and absorb the productivity cost of switching (or minimize that cost), these big companies don't want to be constantly optimizing compute vs rental rate. They want to convert an unruly value (programmer productivity) to something easy, not replace it with a scheduling or optimization problem.
That was working ok until Claude, specifically Claude Code showed up. This was a really useful code-writing harness (that also signed your commits, advertising itself to everyone) that took what are essentially very similar models and made Opus feel like the future of software while GPT 5.2 and friends are just code agents. The performance, ability to handle long term tasks, all of that was basically similar but the harness oriented the model to reason, shell out sub-agents, write scratch code, add console logs, all the sorts of things that 1. seem like science fiction, and 2. improve output a little. Then from fall of last year to no you don't have developers saying "look what I made with LLMs" or "Look what I made with AI" but "Look what I did with Claude". There are not very many blog posts out there about the future of software being re-written due to GPT 5.2 getting autocompaction, but that same feature spawned thousands of "oh shit!" posts in Claude.
That's not a more defensible moat than name recognition + small N for customers. It's a scarier position because if someone else figures out how to deliver the same result (Opus + sonnet + Haiku in a managed ensemble) in a way that was sharp and viral, the same thing they did to OpenAI could happen to them. They still supply the compute but the fact that anyone gives a shit about them is their harness makes it look like more and better code is being written. If that's your situation, you gently write the OpenClaw guy, you threaten to cut off and sue OpenCode for using subscription sign-in. You don't do those things because of a numerator/denominator problem with token cost and monthly fees. You do it because someone using your models in a better harness is a clear brand threat.
Theory 1: the internet has been fully strip mined for all content and is now dead. See that graph of StackOverflow questions dropping off a cliff to zero. Nothing much worthwhile is being added.
Theory 2: they are all unethical as fuck and definitely learning off your data. You would be insane not to - theory 1 means all your free training data is gone, but all that corporate data is fresh, ripe and covers many domains that the amateurs on the internet never filled. You have to launder it some way of course, but it's definitely happening.
Theory 3: winner takes all. I don't care for "Claude" and your wishy-washy ethics performance. ChudAI has a better model and harness? I'm gone this evening.
Having all the users, even if they are exploiting you for cheap compute with their own harness, is essential.
Good theory and insight. Seems like that’s setting us up for some epic big co vs ai co legal battles for covertly training off sensitive and internal big co data
Why would they have that feature in claude code cli if it goes against the ToS? You can use Claude Code programatically. This is not the issue. The issue is that Anthropic wants to lock you in within their dev ecosystem (like Apple does). Simple as that.
allowed shell pipes doesn't necessarily mean they want loops running them.
One of the economic tuning features of an LLM is to nudge the LLM into reaching conclusions and spending the tokens you want it to spend for the question.
presumably everyone running a form of ralph loop against every single workload is a doomsday situation for LLM providers.
> allowed shell pipes doesn't necessarily mean they want loops running them.
insane that people apologize for this at all. we went from FOSS software being standard to a proprietary cli/tui using proprietary models behind a subscription. how quickly we give our freedom away.
Anthropic itself advertised their own implementation of agentic loop (Ralph plugin). Sure, it worked via their official plugin, but the end result for Anthropic would be the same.
There's nothing in TOS that prevents you from running agentic loops.
I don't know why this is downvoted, see my nephew (?) comment [0] for a longer version, but this is not at all clear IMHO. I'm not sure if a "claude -p" on a cron is allowed or not with my subscription, if I run it on another server is it? Can I parse the output of claude (JSON) and have another "claude -p" instance work on the response? It's only a hop, skip, and a jump over to OpenClaw it seems, which is _not_ allowed. But at what point did we cross the line?
It feels like the only safe thing to do is use Claude Code, which, thankfully, I find tolerable, but unfortunate.
Or can you? It's my understanding that you cannot use your subscription with the Agent SDK, that's what the docs say:
> Unless previously approved, Anthropic does not allow third party developers to offer claude.ai login or rate limits for their products, including agents built on the Claude Agent SDK. Please use the API key authentication methods described in this document instead.
Though there was that tweet [0] a while back by someone from Anthropic that just muddied the water. It's frustrating because I feel like the line between the Agent SDK and `claude -p` is not that large but one can use the subscription and one can't... or we don't know, the docs seem unambiguous but the tweet confuses things and you can find many people online saying you can, or you can't.
I'd love to play around with the Agent SDK and try out some automations but it seems I can only do that if I'm willing to pay for tokens, even though I could use Claude Code to write the code "for" the Agent SDK, but not "run" the Agent SDK.
Where is the line? Agent SDK is not allowed with subscription, but if I write a harness around passing data to and parsing the JSON response from `claude -p '<Your Prompt>' --output-format json` would that be allowed? If I run it on a cron locally? I literally have no idea and, not wanting my account to be banned, I'm not interested in finding out. I wish they would clarify it.
You can easily automate OpenCode - more so than the basic Claude Code or Claude desktop app - in a way that automatically uses the maximum amount of subscription quota every cycle. And in an inefficient way that Anthropic can't cache on their end.
If you know anything about subscription models, you know that ALL of them are built on the fact that most of the users don't use the full capacity available all the time.
1. If alternate UX exists, the user can easily replace them with another model as soon as it comes out.
2. All 'all-you-can-eat' plans everywhere comes with clause. Whether it is lunch at a restaurant or it is token-proxy-providers who might think of reselling Max plan to individuals at 20% markup.
> Let the weirdos at Anthropic do what weirdos do, and hopefully one day their name is never mentioned again in polite society.
Woof, that is a bit harsh... :) OpenAI will also face the same problem. They are doing it right now because they need to stand out in some way.
I disagree that this path is inevitable for any service provider. Case in point: Google Drive offers substantially better per-terabyte pricing for storage than Google Cloud Storage; like 100x+ cheaper. Yet, Google Drive has a beautiful API that is used by everything from the Google Drive desktop client to rclone. This is how most apps work, and this frame of thinking about the internet has worked for 40 years because of course 95% of users are going to use the frontend the company makes for their backend; but that 5% of users left on the fringe are oftentimes the most valuable, they're the ones that are going to pay Google an extra $1000/mo for 50TB of storage, and as long as the internal unit economics are good to go, Google should want that. Less edge-cases they have to deal with on drive.google.com, more revenue, all good things.
I do fully expect the limits on these subscriptions to be brought down. But that's not the problem people have with Anthropic today, nor the problem we'd have with OpenAI when they have to eventually do it. That's just the way of things.
The problem is: These actions by Anthropic scream: "Our internal unit economics are going nuclear and we need to do anything we can to regain control."
Low-key: I think the DoW situation was an inflection point for their usage internally. It spiked up hard after that. Dario spent all of 2025 being told "you're not investing enough into compute", but really didn't listen because he wanted to be "responsible" or whatever, and now they're shopping to every provider trying to find compute and are being told that there isn't any.
I don't think it's any of that. It's plain as day to see that as the owners of the API they can see every request, collect every metric they need. The routing can easily be done in the API gateway layer and doesn't depend on the harness at all. Good harnesses are not performing acrobatics around user behaviour to improve anything. They just understand the provider API very well and translate tool calls really well. Then they get out of the way to let the magic happen.
This is about controlling user behaviour, resisting standardisation efforts, to keep subscribers hooked to their models. If they keep their users on claude code, they don't have to adapt to evolving harness standards when shown increased usage across different harnesses. They will try to keep everything on their terms, on their poor quality of engineering, and the users accustomed to claude behaviour will be left with no quick way to switch to other models.
This strategy will keep working as long as they own a top tier frontier model with enough marketing hype behind it to blind unsophisticated users.
My guess is that the telemetry data they can collect from interacting with claude code is the "secret sauce" behind a lot of the improvements we're seeing with coding models right now. Look at cursors Composer-2 release today. Clicking "accept" during plan mode, committing changes and pushing to a remote repo, etc. is a really strong reward signal.
Can't collect telemetry from applications you don't control.
Yup, agreed with this as well. Probably also why they've been investing so heavily in the desktop Claude Code experience; very hard to gather great telemetry from a terminal app.
Literally everyone is desperately trying to figure out why it's so bad and how to make it work consistently using harness etc. But in spite of this massive effort things always go awry after a while. Maybe in a year or two someone figures it out.
This is my theory. They don't want other harnesses to use this because it costs them more. I don't know exactly how OpenCode works, but I'm assuming when people are using this plugin they are mostly using Opus for everything while Claude Code really only uses Opus for writing the actual code. It uses Haiku and Sonnet for almost all of the tasks outside of writing code.
So it hard for them to control and understand the costs of subscriptions if people are using them on different hardnesses that do things that they have no control over.
Yeah, but that's just the model the main agent uses. The subagents aren't Opus. They are Haiku and Sonnet. Most of the token heavy work is offloaded to subagents because of this.
> they must be losing so much money on each Claude Code subscriber that if a million people all said "we're switching" they just wouldn't care.
You're looking at it completely wrong. Claude Code is Anthropic's flagship product, not the API. They want to attract as many users as possible to Claude Code and lock them into their ecosystem, so they can squeeze them later. All of their questionable actions surrounding Claude Code and its subscription are ultimately in service of this goal.
The subscription isn't some kind of charity, it exists specifically because they know the average user isn't willing to pay the exorbitant API prices to vibe code their groundbreaking new B2B SaaS idea, but they want to capture that market share anyway, because it's the core of their long-term strategy. The subscription arose from that: it's a form of predatory pricing designed to attract as many users as possible while they still have VC money to burn.
Once that runs out and the time comes to IPO and start making real profits, they are going to increase the price drastically, and what's where the lock-in comes into play. If everyone is using some open-source alternative that natively supports every other provider on earth, they will be far less likely to continue paying for Claude specifically instead of just switching to a competitor. Not to mention, they'd also lose out on the free advertising from things like CLAUDE.md and the commit co-signing (because that's all those things are, the only reason Claude Code doesn't support AGENTS.md is because CLAUDE.md serves as an advertisement in public repositories).
> like they all think they're Gods above mere mortals, serving some higher purpose, and nothing matters to them except Building the Machine God.
This is all just part of their marketing strategy, and you shouldn't read too much into it.
Anthropic doesn't care and it is all to look good for their IPO.
They are still losing billions of dollars and will do anything to keep people hooked onto the API and will litigate against their own customers.
They will even lobby against open-weight models which is their biggest threat and want to make them illegal to run in the US just for them to succeed.
Anthropic are not your friends and want you to become addicted / over-reliant on Claude Code (hence the free $20 spins at the roulette until March, 27 2026) and charging others on their overpriced API.
Yes, it is true that companies often litigate against customers who violate their Terms of Service. The TOS is put into place to protect the company’s interests from user abuse.
Paying customers of Claude Code don’t receive a free-use license for any desired application. They’re paying to use Claude Code. Anthropic can take steps to litigate usage outside of those terms, even if customers find that fact really annoying.
> I'm really struggling to understand how Anthropic is benefited by not allowing this. Its bad PR for no good reason. The only thing I can figure is that Claude Code is hemorrhaging money, they're too afraid to actually enforce reasonable token limits, and the only thing that's keeping it from totally bankrupting the company tomorrow is: controlling the harness and having the harness dynamically route toward Haiku or Sonnet over Opus when Opus is overloaded, without telling the user.
I have noticed odd behavior when choosing a model, it automatically switches after sometime. Some tasks do not require lots of power, so when I select Haiku, few prompts later, I see Sonnet popping up in the cost/spending. Happened few times.
They're trying for the vertical integration monopoly.
The times it works, it works well for the company at great cost to society.
Imagine the world we'd have if comcast got to control your web browsing experience.
If ISPs got started today, they'd sell the open web at API prices that no one can afford. Then sell the ISP's lock-in 'internet' for a low monthly fee.
My question is why people who don't want comcast's internet think other vertical integrated lock-in is fine.
Our markets game only works for the benefit of society if we have fair markets.
VC-backed loss-leader dumping to starve competition model breaks the game.
They want lock-in for their UI/X, presumably. If Photoshop ran in the cloud, I doubt Adobe would let you make an alternative front-end either. Not that I'm sympathetic to them.
Yes, this. They need as much lock-in as possible before IPO. Most likely less about cash flow and more about IPO story telling.
We'll know for sure when they add full OpenClaw-like features to Claude Code like remote channels & heartbeat support. Both are partially implemented already.
It's served from the cloud but it runs entirely on your PC (except the AI generative tools). It can't run entirely offline though, because the js, webassembly, and assets are served chunked as-needed.
Saying Photopea is good enough is really underselling it. It's so far ahead of anything offered by the open-source community.
I'd be so happy to pay for a fully offline version of Photopea!
Agreed. What I suspect is: the dynamic model routing on CC is way stronger than people realize, and that "Percent-based usage" is intentionally vague because while it is probably measuring "200M tokens per week" or something, they don't want you asking questions about whether you're getting 200M Haiku tokens or 200M Opus tokens. A token is a token to the usage limit, where it comes from doesn't matter to the usage limit. But, to OpenCode it might, because OpenCode can just fire-and-forget everything at Opus (and probably does).
There are a plethora of models that you can use with open code. Anthropic is well within its rights to limit third-party usage of services that violate the TOS. As a Claude code user I’d much rather have the very best experience on Claude code than the largest supportability matrix for Anthropic models.
as someone who has used codex/claude-code/opencode I can confidently say that "the very best experience on Claude" is not the one that Anthropic provides software for.
Well, the challenge for Anthropic's users is: While Anthropic has fantastic coding models right now, they have a bad harness in claude code and the claude desktop app. The best experience using Anthropic's models is OpenCode and Cursor. This wasn't true ~3 months ago, and may return to being untrue in ~3 months, that's how fast these things change, but right now this is the case. Unfortunately, Anthropic models in OpenCode/Cursor are tremendously expensive; and that gets at some of the leading theories on why CC has degraded recently; that Anthropic has been forced to dynamically route more of the agentic process onto Sonnet/Haiku, or reduce the Opus thinking budget.
For all of these reasons, currently, the meta is ChatGPT subscription on Codex or OpenCode. But, again, these things change every few weeks.
I don't think this is as clean-cut as just saying "Anthropic is in their rights" etc. Of course, they are, to whatever degree they are; the bigger problem is that you've got $100/mo and $200/mo Claude subscriptions who are actively saying "the $20/mo Codex subscription is better in every way", possibly because of these thinking budget/routing changes people suspect have happened this month. In other words: Anthropic is at-capacity after the DoW incident, they need to load shed, and they've chosen to harm their high-paying power users and Enterprise over just temporarily slowing growth a bit by hurting the $20/mo plan. And, frankly, they might be right: because for every $200/mo user that jumps over to ChatGPT, half of them will be back once Anthropic can scale capacity, and if they can gain 20x $20/mo users who only use half their sub, that's a win.
> OpenCode is awesome. Claude Code is nothing special at all.
It's weird, my experience is the exact opposite: OpenCode had the weirdest issue of interpreting C:\Users\some-user\Documents\Projects\... as a bunch of escaped characters without \ and created some trash files even with Anthropic models. Before then, OpenCode had the issue of me copying multiple lines of text into it from clipboard somehow quitting the program (crashing?) and then making every following line (of text, notes etc.) be passed to the shell directly. Happened in both Terminal app and also Git Bash on Windows. It might just be that Windows is a shit OS, but Claude Code doesn't really seem to exhibit those problems (and neither does Codex, in their defense). I hope OpenCode has fixed at least some of those since.
> Let the weirdos at Anthropic do what weirdos do, and hopefully one day their name is never mentioned again in polite society.
What an odd thing to say. Maybe I'm out of the loop about some drama (aside from Anthropic having a bit of a spine in regards to govt. while being otherwise flawed still), but pretty happy with my Max subscription, they even fixed the Code GUI mode in their Electron app recently, it finally feels usable and I don't just need to juggle 4 terminal sessions.
This seems harsh and unfair. You aren’t allowed to stream Netflix through a third party streaming service because then the experience is controlled by the third party and there is no lock in to using Netflix and thus no benefit to the subsidy they give you in the plan.
OpenAI is allowing it as a PR stunt and because they have seemingly unlimited cash they can throw at user growth.
One possible factor is reinforcement training: by forcing users to use their own CLI, they gather usage and training data which serves as feedback training data for their own tools. Users operating with a different agent with different tool implementation and quirks might poison that data.
The sibling reply given by strictnein is very likely a factor.
Since I discovered pi I cancelled my Claude subscription and subscribed to ChatGPT. On one hand, the competition is making miracles. On the other hand, it's pretty dooming that there is only one (1) company that keeps my agent cost reasonable.
Extensible coding agent written in typescript. It’s exactly what you (I’m projecting) want out of Claude Code if you’re okay investing time into building your harness or prompting an agent to build it.
The business driver I assume is not lock-in on UX (as some say here) but the additional signal Anthropic gets when using their harness vs a 3P one. It makes sense to discount the price if that signal helps you improve your models, but that discount makes no sense if the user is running your model in another harness and you just get regular API usage signal.
You are wayy overestimating the negative press. HN commenters are a negligible fraction of Anthropic's user base, and I can guarantee that even people here will forget about this and go right back to using Claude Code in a couple days when there is something else to be outraged about. The company needs to do what is best for its business.
reply